2475-1472 (c) 2018 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information. This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/LSENS.2019.2909119, IEEE Sensors Letters 1 Robust Approach based on Convolutional Neural Networks for Identification of Focal EEG Signals Varun Bajaj 1 , Member, IEEE, Sachin Taran 1 , Erkan Tanyildizi 2 , and Abdulkadir Sengur 2 Abstract—Electroencephalogram (EEG) signals provide im- portant information for identification of epileptogenic area. Identification of focal EEG signals locates the epileptogenic area which is an important task for successful surgery. In this paper, convolutional neural networks (CNNs) based framework is proposed for automatic identification of focal EEG signals. The proposed intelligent system initially uses a window for randomly segment the input EEG signals. The short time Fourier transform (STFT) is applied on to the segmented EEG signals for conversion of input EEG signals into Time-Frequency (T-F) representation. The T-F representation of EEG signals are used as T-F images. Instead of training an end-to-end CNNs which necessitates more input images and time, we opt to use a pre-trained CNNs model for transfer learning. Specifically, the deep feature extraction is employed for acquiring the more convenient features from the input EEG images. The deep features extracted from AlexNet, VGG16, VGG19, and Resnet50 models used as input to different variant of k-nearest neighbor (k-NN) classifier. The conducted experimental works show that AlexNet, VGG16, and Resnet50 achieve promising results. Specifically, the fc6 layers of AlexNet, VGG16 and fc1000 layer of Resnet50 produce 99.8% accuracy score with weighted-k-NN approache. The comparative study shows that proposed method provides better performance in comparison to state-of-the-art methods. Index Terms—Electroencephalogram (EEG) signal; convolu- tional neural networks (CNNs); transfer learning; k-nearest neighbor . I. I NTRODUCTION T HE electroencephalogram (EEG) signal changes with electrical activity of brain. The EEG signal is the valuable clinical tool for investigating the pathological states of the human brain such as epileptic seizure. Epilepsy is unexpected electrical disturbance of brain. The focal epilepsy is the pathological state of brain which affects either the entire hemisphere or limited portion of the hemisphere. The affected portion hemisphere recorded EEG signals are referred to focal class (FC) and not affected portion hemisphere signals are corresponds to non-focal class (NFC) [1]. For FC EEG signal identification, the traditional visual inspection methods are limitedly used due to time consuming process and requirement of trained clinician. Therefore, it is much needed to develop an automatic method that can provides an accurate solution for the classification of FC and NFC EEG signals. The mean and standard deviation of Euclidean distances are used for discrimination of FC and NFC EEG signals [2]. Delay permutation entropy based features (DPE) [3], average sample entropy and variation of instantaneous frequencies of IMFs [4], features extracted from discrete wavelet transform (DWT) [5], The authors are with the Electronics and Communication Discipline, Indian Institute of Information Technology Design and Manufacturing, Jabalpur, 482005 MP, India 1 . Firat University, Technology Faculty, Electrical and Elec- tronics Engineering Dept., Elazig, Turkey 2 . E-mail: ( varunb@iiitdmj.ac.in; taransachin2@gmail.com; etanyildizi@firat.edu.tr; sengur@gmail.com) feature average bandwidths ratio (AvgBratio) extracted from AIMFs [6], and the entropy features from empirical mode decomposition (EMD)-DWT domain are used for classification of FC and NFC EEG signals [7]. Features based on dual tree complex wavelet transform (DT- CWT) [8], flexible analytic wavelet transform (FAWT) [9], and tunable-Q wavelet transform (TQWT) are used for identifica- tion of FC EEG signals [10]. The nonlinear features [11] and entropies in wavelet domain [12] are used as features for FC EEG signals identification. Rhythms based features [13] and area parameters computed from two-dimensional projections of the reconstructed phase space are introduced for FC and NFC EEG signals classification [14]. Bivariate-EMD explores the complexity of EEG signals for focal activity identification [15]. Nonlinear features are explored for identification of FC EEG signals [16]. The clustering- variational mode decom- position (CVMD) based spectral features are introduced for FC EEG signals identification [17]. The features from time, frequency and statistical measurements have been used as input to SVM classifier for classification of focal and non-focal EEG signals [18]. In this work, CNNs based deep features are explored for identification of focal EEG signals. The STFT based time frequncy (T-F) images used for resize and deep features extraction, which further fed into k-NN classifier for identification of focal EEG signals. The paper is organized as follows: section II consist of methods and Materials, section III defines the description of proposed methods, section IV provides the detailed results and discussion, and finally section V defines the conclusions of the proposed work. II. METHODS AND MATERIALS A. Dataset The focal EEG-database could be find online at www.dtic.upf.edu/ralph/sc/ [1]. The each EEG signal contains signals x and y recorded by adjacent channel, the difference of recording is used for processing, with sampling frequency is 512 Hz. The recorded signals are preprocessed by a band- pass Butterworth filter, which provided the frequency content in range of 0.5 Hz to 150 Hz. In this database, the each FC and NFC contain 3750 number of EEG observations. B. Convolutional Neural Networks (CNNs) Convolutional neural networks (CNNs) are the category of deep, feed-forward artificial neural networks (ANNs), which find application for analysis and classification of images [19][20][21]. The end-to-end learning architecture of CNNs provides attribute extraction and classification. As compared to traditional single layer ANNs architecture, CNNs consist of multiple layers as convolution, pooling, and fully connected